Inter-American Development Bank Banco Interamericano de Desarrollo (BID) Research Department Departamento de Investigación Working Paper #554 The Unexplained Part of Public Debt by Camila F.S. Campos* Dany Jaimovich** Ugo Panizza** *Yale University **Inter-American Development Bank, Washington, D.C. March 2006
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Inter-American Development Bank
Banco Interamericano de Desarrollo (BID) Research Department
Departamento de Investigación Working Paper #554
The Unexplained Part of Public Debt
by
Camila F.S. Campos* Dany Jaimovich**
Ugo Panizza**
*Yale University **Inter-American Development Bank, Washington, D.C.
March 2006
2
Cataloging-in-Publication data provided by the Inter-American Development Bank Felipe Herrera Library Campos, Camila F.S.
The unexplained part of public debt / by Camila F.S. Campos, Dany Jaimovich, Ugo
Panizza.
p. cm. (Research Department working paper series ; 554) Includes bibliographical references. 1. Debts, Public. 2. Budget deficits. 3. Financial statements. I. Jaimovich, Dany. II.
Panizza, Ugo. III. Inter-American Development Bank. Research Dept. IV. Title. V. Series.
This paper shows that budget deficits account for a relatively small fraction of debt growth and that stock-flow reconciliation, which is often considered a residual entity, is one of the key determinants of debt dynamics. After having explained the importance of the stock-flow reconciliation, the paper shows that this residual entity can be partly explained by contingent liabilities and balance-sheet effects.
1 The views expressed in this paper are the authors and do not necessarily reflect those of the Inter-American Development Bank. The usual caveats apply. Camila Campos: [email protected], Dany Jaimovich: [email protected], Ugo Panizza: [email protected].
4
1. Introduction How do countries get into debt? The answer to this question may seem trivial. Countries
accumulate debt whenever they run a budget deficit (i.e., whenever public expenditure is higher
than revenues). In fact, the standard Economics 101 debt accumulation equation states that the
change in the stock of debt is equal to the budget deficit:
ttt DEFICITDEBTDEBT =− −1 (1)
and that the stock of debt is equal to the sum of past budget deficits: ∑=
−=t
iitt DEFICITDEBT
0.
Whoever has worked with actual debt and deficit data knows that Equation (1) rarely holds and
that debt accumulation can be better described as:
tttt SFDEFICITDEBTDEBT +=− −1 (2)
where tSF is what is usually called “stock-flow reconciliation.” Clearly, Equation (1) is a good
approximation of debt accumulation only if one assumes that tSF is not very large. The purpose
of this paper is to describe some of tSF ’s main characteristics. The paper shows that, contrary to
what is usually assumed, the budget deficit accounts for a small fraction of the within-country
variance of the change in debt over GDP and that stock-flow reconciliation plays an important
role in explaining debt dynamics. The paper also shows that, on average, tSF tends to be positive
and that there are large cross-country differences in the magnitude of this residual entity. This
suggests that the magnitude of stock-flow reconciliation is not likely to be purely due to random
measurement error. In particular, the paper shows that the problem is especially serious in
developing countries and, among this group of countries, the difference between debt and deficit
is particularly large in Latin America and Sub-Saharan Africa.
The paper also runs a set of regressions aimed at explaining the main determinants of the
magnitude of the stock-flow reconciliation and finds that balance-sheet effects due to real
depreciations and contingent liabilities that arise at time of banking crises are strongly correlated
with the difference between deficit and change in debt. However, the paper also shows that the
regressions can only explain 20 percent of the within-country variance of the stock-flow
5
reconciliation and that there is still much that we do not understand about one of the main
determinants of debt accumulation.
While we are not the first to show that stock-flow reconciliation is an important part of
debt dynamic (see, among others IMF, 2003; Martner and Tromben, 2004; European
Commission, 2005; Budina and Fiess, 2005), we are not aware of any other paper that
systematically describes the main characteristics of this residual, but extremely important,
determinant of debt accumulation.
The rest of the paper is organized as follows. Section 2 describes our main sources of
data and presents some basic facts on public debt and deficit. Section 3 focuses on a detailed
description of the stock-flow reconciliation. Section 4 runs a set of regressions aimed at
explaining the main determinants of the stock flow reconciliation. Section 5 concludes.
2. Data
The purpose of this section is to describe our data on fiscal deficit and public debt. In this
context, it is worth mentioning that obtaining reliable and comparable data on the stock public
debt is a rather difficult exercise. In fact, the IMF International Financial Statistics (IFS) and
IMF Government Finance Statistics (GFS), which are the most common sources of cross-country
data on government statistics, report data for a rather limited set of countries. This is even the
case for industrial countries; these sources do not report recent data on public debt for Japan and
Italy, for example. Furthermore, most cross-country datasets do not make an effort to make the
data comparable across countries (for a discussion of these issues, see IMF, 2003).2
Although there are now some papers that attempt to build comparable cross-country data-
sets on public debt (Cowan et al., 2005; Jeanne and Guscina, 2006; IMF, 2003; Budina and
Fiess, 2005), some of these data sets are not publicly available and all of them have a limited
country and time coverage. As a consequence, we do not rely on these new data and only use
publicly available sources (hence, the caveats mentioned above should be kept in mind). In
particular, we start with IFS and GFS and supplement them with data collected from national
sources (mostly from the websites or publications of the various Ministries of Finance), the UN
Economic Commission for Latin America and Caribbean (ECLAC, see Martner and Tromben,
2004), and the Organization for Economic Cooperation and Development (OECD). 2 The most important problems include the treatment of sub-national governments and the use of gross versus net debt (for a methodological note, see Cowan et al., 2005).
6
Using these various sources, we assemble an unbalanced panel covering 117 countries
and consisting of approximately 1,900 observations. Table A1 in the Appendix lists all the
countries included in our dataset, the time coverage for each country, and summary statistics for
debt and deficit ratios. Our sample includes 24 high-income countries, 59 middle-income
countries and 34 low-income countries. The regions with the largest number of countries are
Sub-Saharan Africa (27 countries) and Latin America (25 countries). South Asia and East Asia
are the regions with the smallest number of countries (five and eight countries, respectively).
While long time series are available for some countries (e.g., Bahamas, Burundi, Costa Rica,
Iceland, Norway and the US have more than 30 years of data), for others there are very few
observations (Albania, Algeria, Gabon, Sudan, Togo, and Yemen are among the countries with
less than five years of data).
Table 1 shows that the sample mean of the deficit to GDP ratio is 4.04 percent and that
average deficit tends to decrease with the level of income. The region with the highest average
deficit is South Asia (6.5 percent), followed by the Middle East (5.6 percent), and Sub-Saharan
Africa (4.2 percent). Latin American countries tend to have fairly low levels of average deficit
(just below the cross-country average) but the region is far from being homogeneous and is
characterized by the largest variance in the sample.
Table 2 reports summary statistics for the debt-to-GDP ratio and shows that the cross-
country average is close to 56 percent. South Asia and Sub-Saharan Africa are the regions with
the highest levels of debt (67 and 60 percent, respectively) and East Asia and Eastern Europe and
Central Asia are the regions with the lowest level of debt (35 and 37 percent, respectively). Latin
America has a level of debt that is just below the sample average and is not much higher than
that of the industrial countries included in our sample. Again, we find that Latin America is one
of the most heterogeneous regions in our sample (in this case, second only to Sub-Saharan
Africa). As one may expect, we find that most of the variance in debt-to-GDP is due to
differences across countries (this is the between standard deviation). However, there is also
substantial variance within countries. In fact, the within standard deviation (not reported in the
table) is often close to 50 percent of the between standard deviation.
7
Table 3 focuses on the change in debt divided by GDP ( tid , ).3 If Equation (1) were to
hold, the change in debt should be equal to the budget deficit. By comparing Table 2 with Table
3, we find that the value of tid , is almost five percentage points higher than average deficit over
GDP, indicating that more than 50 percent of the average change in debt is not explained by
deficit.4 The Table also shows that while the difference between tid , and the deficit is fairly small
in industrial countries (about 0.3 percentage points), this difference is extremely large in Latin
America and Sub-Saharan Africa, where the average deficit is about one-third the average
change in debt.
We can now describe the characteristics of the stock-flow reconciliation by defining the
following measure of the difference between change in debt and deficit for country i at time t.
( )100
,
,1,,, ×
−−= −
ti
titititi Y
DEFICITDEBTDEBTδ (3)
Clearly, ti ,δ is just the stock-flow reconciliation of Equation (1) expressed in terms of
GDP (ti
ti
ti Y
SF
,
,
, =δ ). Table 4 describes ti ,δ and shows that the change in debt is nearly five
percentage points higher than the deficit (with the highest values in Latin America and Sub-
Saharan Africa). However, the Table also shows that there are several countries with extremely
large values of ti ,δ (in some cases well above 200 percent). In Latin America, for instance, the
difference between the change in debt and deficit has a range of 350 percentage points (from –73
3 It is important to note that we do not use the change in the debt-over-GDP ratio (i.e., 1001
1, ×⎟⎟
⎠
⎞⎜⎜⎝
⎛−=
−
−
t
t
t
tti Y
DYDθ )
but the change in debt divided by GDP at time t (i.e., 100)1(1
1, ×⎟⎟
⎠
⎞⎜⎜⎝
⎛+
−=−
−
gYD
YD
dt
t
t
tti ). As nominal GDP
growth (g) tends to be positive, tid , is usually larger than ti,θ . We use this measure, rather than the standard ti,θ because we want to isolate changes in debt from changes in the level of GDP. 4 Using a different methodology and a shorter sample, IMF (2003) also finds similar but less drastic results. In particular, it finds that more than 25 percent of the increase in the debt-to-GDP ratio of a sample of emerging market countries over the 1997-2003 period is due to off balance-sheet factors. In a sample of 21 market-access countries, Budina and Fiess (2005) find that debt over GDP increased by 22.8 percentage points from 1994 to 2002, while real GDP grew by 9.3 percent, yielding a change in debt of approximately 37 percent. The deficit (primary plus interest rate bill) explained about one-third of this change while other factors (including the real exchange rate) explained the remaining two-thirds.
8
to 281). The industrial countries have the smallest range, but even in this case the range is close
to 30 percentage points. These extreme values are due either to exceptional events or
measurement error. In the second column of Table 5, the average value of ti ,δ is computed by
dropping the top and bottom 2 percent of the distribution. After dropping these outliers, we find
that ti ,δ has an average value of 3 percent and that the average values of ti ,δ for Latin America
and the Middle East drop from 7 percent to 4 and 2 percent, respectively.
It is also interesting to see which countries tend to have large values of ti ,δ . Table 5
summarizes all the episodes for which 10, >tiδ (a full list of episodes is reported in Tables A2
and A3 in the appendix). There are 238 country-years (corresponding to 13 percent of
observations) for which 10, >tiδ , and 50 country-years (3 percent of observations) for
which 10, −<tiδ . The industrial countries, East Asia, and South Asia are the regions with the
lowest number of episodes (and very few episodes where 10, −<tiδ ). Sub Saharan Africa, the
Middle East and North Africa, and Latin America are the regions with the largest number of
episodes.
While this paper focuses on change in debt, we obtain the same results if we use the
standard decomposition of the change in debt over GDP (θ).5 Figure 1 shows that in most regions
the stock flow adjustment is the main determinant of debt growth and inflation is the main
determinant of debt reduction
3. Debt and Deficit
The previous section showed that simple comparisons of average values of deficit over GDP and
change in debt indicate that Equation (1) is far from being a good approximation of the main
determinants of debt accumulation and that what is usually considered a residual entity (the
5 The standard decomposition takes the following form:
( )t
t
t
t
t
t
t
t
t
t
t
t
YSF
gYDEBT
grgY
DEBTi
YPD
YDEBT
YDEBT
++
+−+
+=−−
−
−
−
−
−
)1()1( 1
1
1
1
1
1 π
where the first term on the RHS of the equation is the contribution of the primary deficit, the second term is the interest bill, the third term is the contribution of nominal growth (which can be split into real growth and inflation) and the last term is the stock-flow adjustment.
9
stock-flow reconciliation) is a key determinant of debt accumulation. In this section, we use
different strategies to provide more evidence in this direction.
3.1 Regressions Analysis
One way to assess the importance of tSF is to divide debt and deficit by current GDP and use
our large panel to estimate the following fixed effects regression:
ititiit defd ,,, * εβα ++= (4)
where iα is a country fixed effect (the country fixed effects control for the fact that the data
come from different sources, countries have different levels of debt, and they use different
methodologies for computing debt and deficit) and itdef , is deficit over GDP. If Equation (1)
holds, we expect a high R2 (the regression’s R2 should be 1 if Equation 1 holds exactly), iα =0,
and β =1. Hence, the regression’s coefficients and R2 can be used to asses the relative
(un)importance of the deficit in explaining changes in debt. Table 6 reports the results of the
estimation of Equation (4) for different sub-samples of countries. Column 1 describes the basic
pattern. First of all, we find that β is greater than 1 (but not significantly different from 1)
indicating that a 1 percent increase in the deficit to GDP ratio tends to translate into a 1.3 percent
increase of the debt to GDP ratio. More interestingly, the regression’s R2 shows that, in our
sample of countries, deficits explain less than 8 percent of the within country variance of itd , and
that tSF explains more than 90 percent of the variance.6
As the low R2 could be due to the presence of outliers, in Column 2 we drop 47 outliers
(defined as observations that have residuals with an absolute value greater than 2.5 standard
deviations). After dropping these outliers, β drops to 1.18, but we still find that our model can
only explain 23 percent of the variance of itd , . Figure 2 plots the fit of the regression reported in
Column 2 and illustrates that the low R2 is not due to a few episodes with a particularly low fit,
but that most countries have observations that are far away from the regression’s line. Column 3
6 We also ran separate regressions for the 58 countries for which there are at least 15 years of data. We found that β had average and median values of approximately 1 and ranged between –1.8 (Zaire) and 5.9 (Rwanda). The regressions’ R2 had an average value of 0.32, a median value of 0.25, and ranged between 0.007 (Egypt) and 0.87 (Italy). There are only four countries (all industrial) that have an R2 above 0.8, 16 countries (11 of them industrial) for which the R2 is higher than 0.5, and 18 countries for which the R2 is less than 0.1.
10
of Table 4 addresses the outlier issues by running the same regression as in Column 1 using a
median quantile regression with bootstrapped standard errors (STATA’s BSQREG) and shows
that in this case, the coefficient of the deficit variable drops to 0.87 and the R2 goes to 0.24.
The remaining columns run separate regressions for different regions of the world.
Column 4 focuses on 29 countries located in Sub-Saharan Africa and finds that the deficit
explains only 3 percent of the variance of itd , . Columns 5 and 6 show that in Latin America and
the Caribbean (25 countries) and South Asia (5 countries), the deficit explains between 5 and 6
percent of the variance of itd , . Columns 7 and 8 focus on East Asia (8 countries) and the Middle
East and North Africa (11 countries) and show that the deficit explains between 14 and 20
percent of the within country variance of itd , . The developing region with the best fit is East
Europe and Central Asia (Column 9, 15 countries). In this case, the deficit explains 23 percent of
the variance of itd , . Only in the sub-group of industrial countries (Column 10, 24 countries) does
the deficit explain more than one-quarter of the within country variation of itd , but even in this
case, the regression can only explain half of the variance of the dependent variable.
3.2 Theoretical R2
As an alternative way to describe the pattern documented above, we build a measure aimed at
determining which countries have the largest deviation from the theoretical identity defd = .
Clearly, such a measure cannot be the country average of ti ,δ described in Table 5 because
negative and positive values of ti ,δ would compensate each other. One possibility would be to
adopt a strategy similar to the one of the previous section and run country-by-country regressions
of DEBTΔ over DEFICIT and use the fit of these regressions (their R2) as a measure of how
much a country deviates from defd = . One problem with this strategy is that it would not help
to differentiate countries that have a good fit in which defd = holds, from countries that have a
good fit but where the relationship between debt and deficit can be better described with an
equation of the type: ttt defd εβα ++= * with 0≠α and 1≠β . An index that addresses these
problems and relates to a regression’s R2 can be defined as:
11
( )
( )∑
∑
=
=
−= T
titi
T
tti
i
dd1
2,
1
2,δ
φ (5)
Note that iφ is always non-negative and naturally relates to the R2 of a regression of
tid , over def. In fact, if we write ttt defd εβα ++= * and, if instead of estimating the
regression’s parameter, we force 0=α and 1=β , the R2 of the model would be 1- iφ . Hence, if
the true parameters describing the relationship between debt and deficit were 0=α and 1=β ,
iφ would be equal to 0. Thus, higher values of iφ indicate larger deviations of the true
parameters from 0=α and 1=β . Figure 3 illustrates the theoretical distribution of iφ for
different values of β under the assumptions that 0=α , 10=α , and 10−=α . The figure shows
that when 0=α the distribution is asymmetrical with iφ rapidly going towards infinite when β
tends to 0, and iφ converging to around 1.5 when β goes to infinite, the figure also shows that
iφ is equal to 0 when β =1. When 10=α , the distribution becomes monotone but still going to
infinite when β goes to 0 and converging to approximately 1.5 when β goes to infinite. When
10−=α the distribution reaches a minimum when β is around 4 and then starts increasing and,
again, converges at around 1.5.
Figure 4 shows the values of iφ for our sample of countries. Few countries have a value
of iφ close to 0 and most countries are concentrated in the 0.5-1.5 range. In particular, 15 percent
of countries have values of iφ that are below 0.5 (the lowest value, 0.009, is for Finland), 30
percent of countries have values that range between 0.5 and 1, 35 percent of countries have
values that range between 1 and 1.5, and the remaining 20 percent have higher values. Table 7
shows that the mean and median of the distribution of iφ is approximately 1 and that, as
expected, the industrial countries have the lowest value of iφ and Latin America and the Middle
East have the highest values of iφ .7
7 It may seem surprising that while the theoretical distribution is highly skewed, the data of Table 7 indicate that the mean is identical to the median. This is due to the fact that Table 7 does not include four countries that have values of φ greater than 4 (these countries are Estonia, Seychelles, Luxembourg, and Sudan). If we include these countries, the median goes to 1.05, but the average jumps to 2.7.
12
3.3 Debt Explosions
So far, we documented that there are a large differences between deficit and change in debt. Now
we explore whether the difference between these two variables is positively correlated with debt
growth. Figure 5 plots the relationship between the growth rate of debt over GDP (defined as
( ) 1001,1,,,, ×−= −− tititititi YDYDθ ) and the ratio between deficit and change in debt (defined as
tititi ddef ,,, =ρ ).8 It shows that at relatively low levels of debt growth (below 5 percent per
year), the deficit explains approximately 80 percent of the change of debt. However, when debt
starts growing at a faster rate, the share of debt explained by deficit drops dramatically. In
particular, the figure shows that when annual debt growth reaches 10 percent of GDP, the deficit
explains less than 40 percent of debt growth. Table 8 regresses ti,θ over ti,ρ (controlling for
country fixed effects) and confirms that there is a negative and statistically significant
relationship between these two variables. While the fit of the regression is rather poor, the table
shows that the fit improves if extreme values of ti,θ are not considered (compare, for instance,
Column 1 with Column 3 where episodes in which ti,θ >50 are dropped). The table also shows
that the relationship between ti,θ over ti,ρ does not vary much across groups of countries.
As a last exercise, we look at debt explosions (defined as episodes in which ti,θ >10);
Table 9 summarizes the data and Table A4 lists all the episodes. The first panel of Table 9 shows
that in the 172 episodes for which ti,θ >10 (9 percent of the country-years for which we have
data), the average increase in debt over GDP was close to 28 percentage points, the average
change in debt was around 46 percentage points (the difference between these two values is
nominal GDP growth which, in presence of high inflation, can be very high), and the average
ratio between these two variables was 70 percent. The fourth column of the table shows that in
our sample of debt explosions, average deficit was close to 10 percent of GDP and the ratio
between deficit and change in debt was about 27 percent. This is close to one-third of the same
ratio during normal times (when 10> ti,θ >0 the ratio between deficit and change in debt is 75
percent). The table also shows that the regions with the highest occurrence of debt explosions are
Latin America and Sub-Saharan Africa (41 and 66 episodes, respectively) and that East Europe
8 We smooth the curve with a bandwidth of 25.
13
and Sub-Saharan Africa are the regions with the lowest average ratio between deficit and change
in debt (18 and 13 percent, respectively).
Since the average values discussed above may be driven by extreme values of ti,θ , we
restrict the sample in the second panel of Table 9 to 104 episodes for which ti,θ ranges between
10 and 20 percent. In this case, we find that the average increase of the debt-to-GDP ratio is
approximately 14 percent, the average change in debt is 24 percent and the average ratio between
these two variables is 68 percent (basically identical to the top panel of the table). The fourth
column of the table shows that the average deficit is 7 percent and that the ratio between average
deficit and change in debt is 29 percent, which again is close to the top panel of the table. As
before, we find that Latin America and Sub-Saharan Africa have the highest occurrence of debt
explosions (18 and 36, respectively), but now we find that the Middle East and the industrial
countries have a number of episodes that are not much lower than those of Latin America. In
fact, we now find that Latin America has the second lowest (after the industrial countries)
relative share of debt explosions. This confirms that debt explosions in Latin America tend to be
very large. In fact, Latin America is the only region in the world where there are more episodes
in which debt grows by more than 20 percent of GDP than episodes in which debt grows
between 10 and 20 percent of GDP.
4. What Drives the Difference?
After having documented that there are large differences between deficits and change in debt, we
now run a set of regressions aimed at exploring the determinants of these differences. We start
by estimating the following model:
tititiiti X ,,,, εγπβαδ +++= (6)
where iα is a set of country fixed effects, tiX , a set of country-year specific variables that can
explain the difference between deficit and change in debt, and ti,π is a measure of inflation
(defined as ln(1+INF)). Although we do not have a clear theory of how inflation should affect
ti ,δ , we include this variable because the various components of ti ,δ are nominal variables
measured in different periods of time (a stock at time t, a stock at time t-1 and two flow variables
measured between t-1 and t). Hence, whenever the deficit is different from the change in debt,
14
the value of ti ,δ should be positively correlated with nominal GDP growth, which is heavily
influenced by inflation.
One reason why the change in debt could be higher than the recorded deficit is the
valuation effects due to currency depreciations in the presence of foreign currency debt. To
explore this possibility, we start by focusing on developing countries (industrial countries do not
have large stocks of foreign currency debt) and use data from the World Bank’s Global
Development Finance (GDF) to create three dummy variables that classify all developing
countries into three groups of equal size.9 The three dummies are defined as follows: (i) LOW
takes a value of 1 for all country-years where the external debt-to-GDP ratio is below 38 percent;
(ii) MEDIUM takes a value of 1 for all country-years where the external debt-to-GDP ratio
ranges between 38 and 64 percent; (iii) HIGH takes a value of 1 for all country-years where the
external debt-to-GDP ratio is above 64 percent. Next, we interact the three dummies with the
change in the real exchange rate (DRER, an increase in DRER corresponds to a real
depreciation).
Column 1 of Table 10 reports the results of our baseline estimation. As expected, we find
that inflation has a positive and statistically significant coefficient. Furthermore, we find that
currency depreciations are positively and significantly correlated with δ , a finding that provides
evidence of the presence of balance-sheet effects. More interestingly, we find that the effect of
currency depreciations is particularly large in countries with high levels of external debt.
Consider, for instance, a real depreciation of 30 percent (not an uncommon event in some of the
countries included in our sample). In countries characterized by low or medium levels of external
debt, such a depreciation is associated with an increase of δ of approximately three to four
percentage points, but in countries with high levels of debt, a similar depreciation would instead
cause δ to increase by more than 10 percentage points. At the bottom of the table we show that
the difference between coefficients is also statistically significant (this is not the case for the
difference between the coefficients associated with low and medium external debt).
Next, we include industrial countries and assume that this set of countries has no foreign
currency denominated external debt. Therefore, the regression coefficients should be interpreted
9 Since the GDF data have information for total external debt, we are implicitly assuming that most external debt is public (or generates contingent liabilities of the public sector). We checked the validity of this assumption by computing the correlation between GDF data on total external debt and IFS data on public external debt and found that this correlation is 0.91.
15
as follows: DRER measures the effect of real depreciations in industrial countries;
DRER+DRER*LOW measures the effect of a real depreciation in developing countries with low
levels of external debt; DRER+DRER*MEDIUM measures the effect of a real depreciation in
developing countries with average levels of external debt; and DRER+DRER*HIGH measures
the effect of a real depreciation in developing countries with high levels of external debt. Column
2 shows that the coefficient of DRER is low and not statistically significant, indicating that there
are no balance-sheet effects in industrial countries. As before, we find that balance-sheet effects
are important in developing countries and that the effect of a real depreciation in all three groups
of developing countries is significantly different (both in economic and statistical terms) from
the effect of a depreciation in industrial countries. Finally, we still find that balance-sheet effects
tend to be particularly important in countries with high levels of debt.
Column 3 explores the role of default, w expect defaults to be associated with debt
reduction and hence negatively correlated with δ . To capture the effect of default, we use data
from Standard and Poor’s and build a dummy variable that takes a value of 1 around the last year
of a default episode (in particular, it takes a value of 1 in the last year of the episode and in the
year before and the year after the last year of the episode). Next, we build a default dummy that
takes a value of 1 in the last year of a Paris club rescheduling and then another dummy that takes
a value of 1 whenever the GDF reports that a country has rescheduled its debt. Finally, we build
a dummy called DEFAULT that takes a value of 1 whenever one of the previously described
dummies takes a value of 1. Column 3 shows that the default dummy has the expected negative
sign but that the coefficient is small and not statistically significant (we obtain similar results if
we use the three dummies separately).
Column 4 uses data from Caprio and Klingebiel (2003) to explore the role of banking
crises. These are important events because they generate a series of contingent liabilities and
other off-balance sheet activities that can translate into debt explosions. As expected, we find
that the coefficient of the banking crisis dummy is positive and statistically significant. The
coefficient is also quantitatively important, indicating that the average banking crisis is
associated with an increase of three percentage points in δ .
Column 5 jointly includes all the variables discussed above. We find that the results are
qualitatively similar to previous ones, but that the coefficient of DRER*MEDIUM is no longer
statistically significant (however, DRER+ DRER*MEDIUM remains significant) and that the
16
same is true for banking crisis. In the last column of the table, we control for year fixed effects
(which implicitly control for global shocks) and show that their inclusion does not affect our
basic results.
It is interesting to note that the set of controls included in the regressions of Table 10
explains about 20 percent of the variance of δ and that the country fixed effects explain about
30 percent of the variance of δ (see last row of Table 10). This indicates that country specific
factors explain most of the variance of δ and corroborates the findings of Table 4, which
showed that there are large cross-country differences in the average value of δ . There are two
possible explanations for this finding. The first has to do with the fact that measurement errors
that lead to an underestimation of the deficit are more important in some countries than in others,
which is probably related to the fact that poorer countries have less sophisticated accounting and
budgeting systems. The other has to do with the fact that the importance of contingent liabilities
that lead to debt explosions vary across countries and that our set of controls does not capture all
these contingent liabilities.10
Table 11 includes GDP growth in the analysis. The first column shows that debt tends to
grow more than deficit during periods of slow GDP growth. Column 2 substitutes GDP growth
with two dummies variables that take a value of 1 during periods of high growth (GOOD
TIMES) and periods of slow growth (BAD TIMES).11 Also in this case, we find that debt tends
to grow faster than the deficit during bad times and slower than the deficit during good times.
Column 3 augments the regression in Column 1 with the set of controls in Table 10. We find that
the sign of GDP growth remains negative but the coefficient drops by one-third and is no longer
statistically significant. Column 4 uses the set of controls in Table 10 and the GOOD TIMES and
BAD TIMES dummies. In this case, we still find that the two dummies have the opposite sign
and are both statistically significant.
In Table 12 we estimate a set of regressions similar to those in Table 10 but now
substitute δ with d and include def in the set of controls. This is equivalent to estimating the
model of Table 10 by relaxing the restriction that the coefficient of def is 1. We find that the def
coefficient is always smaller than 1 but that that this coefficient is never significantly different
10 Another key difference is in the size of the regional government, which is often not well captured by our data. 11 GOOD TIMES takes a value of 1 when growth is one standard deviation above the country average, BAD TIMES takes a value of 1 when growth is one standard deviation below the country average. REGULAR TIMES is the excluded dummy.
17
from 1. All our other results are unchanged (this was expected because Table 6 already indicated
that the deficit by itself explains an extremely small share of the within-country variance of the
change in debt).
One problem with the regressions of Tables 10, 11 and 12 is that they assume a linear
relationship between the dependent variable and the set of independent variables. Therefore, the
estimated results might be driven by extreme values of δ . To address this issue, we relax the
linearity assumption and run two sets of Probit regressions. In the first set of Probits, the
dependent variable is a dummy that takes a value of 1 for all country years in the top decile of
the distribution of δ . In the second set of Probits, we repeat the experiment using the bottom
decile of the distribution of δ. 12
Table 13 reports the results for events in the top decile (in this group of events, δ ranges
between 12.7 and 282 and has an average value of 44.5). We find that most of the results are
similar to those in Table 10. In particular, Column 1 shows that the relationship between real
depreciations and the probability of observing an extreme event of δ increases with the level of
external debt. Column 2 shows that in industrial countries, real depreciations have a negative
(but not statistically significant) correlation with the probability of observing an extreme event of
δ. This column also shows that in countries with high levels of external debt, depreciations are
highly correlated with the probability of observing an extreme event. One puzzling result of
Table 13 is that the coefficient of the DEFAULT dummy is large, significant, and positive
(Column 3). This is exactly the opposite of what we expected, and may have to do with the fact
that defaulted debt is not immediately subtracted from the stock of public debt. The coefficient of
the BANKING CRISIS dummy variable instead has the expected positive sign. Besides being
statistically significant, the impact of this variable is also economically important. In particular,
the point estimates indicate that a banking crisis is associated with a 10 percent increase in the
probability of observing an extreme event of δ.
Table 14 focuses on events in the bottom decile of δ (in this group of events, δ ranges
between -116 and –3.4 and has an average value of -10.9). As expected, we find that
depreciations are negatively correlated with these types of events but the coefficients are rarely
significant. In general, we find that our model does a very poor job of explaining these events.
12 The results do not change if we define the dummies using the |δ|>10 threshold.
18
5. Conclusions
The purpose of this paper was to document the fact that what is often considered a residual entity
is indeed one of the key determinants of debt dynamic. After demonstrating the importance of
the stock-flow reconciliation, this paper shows that this residual entity can be partly explained by
contingent liabilities and balance-sheet effects. These results suggest that building a safer debt
structure and implementing policies aimed at avoiding the creation of contingent liabilities are
key to avoiding debt explosions (for contrasting views on how this can be achieved, see
Goldstein and Turner, 2004 and Eichengreen, Hausmann and Panizza, 2003). However, this
paper also shows that a large fraction of the variance of the stock-flow reconciliation cannot be
explained by balance-sheet effects and our simple regressions.13
13 One variable that is likely to be important but that we do not control for is the effect of court decisions that force the government to make payments (to public sector workers, for instance) that were not budgeted. We would like to thank Vito Tanzi for pointing this out.
19
References Budina, N. and N. Fiess. 2005. “Public Debt and its Determinants in Market Access Countries.”
Washington, D.C.: The World Bank.
Caprio, G., and D. Klingebiel. 2003. “Episodes of Systematic and Borderline Financial Crises.”
Washington, DC, United States: World Bank. Mimeographed document.
By Income Groups Low 4.67 4.40 2.76 -18.26 45.15 34 440 Medium 4.13 6.18 4.28 -10.02 66.05 59 947 High 3.29 3.78 2.92 -6.89 20.79 24 485 The income group and regional classifications are those used by the World Bank
Table 2. Debt over GDP
Country Group σ (%)
μ (%)
Overall Between
Min (%)
Max (%)
N. of countries
N. of observations
All Countries 55.80 58.05 46.92 0.00 637.52 117 1872 By Region
By Income Groups Low 72.21 56.50 49.57 1.49 304.50 34 440 Medium 54.27 67.94 48.02 0.00 637.52 59 947 High 43.91 26.75 27.08 1.47 121.53 24 485 The income group and regional classifications are those used by the World Bank. * Excludes Guyana ** Excludes Israel
21
Table 3. Change in Debt over GDP
σ (%) Country Group
μ (%)
Overall Between
Min (%)
Max (%)
N. of countries
N. of observations
All Countries 8.97 23.42 14.66 -118.17 303.57 117 1872 By Region
By Income Groups Low 14.30 31.28 22.25 -118.17 243.68 34 440 Medium 9.00 24.39 11.54 -61.52 303.57 59 947 High 4.05 4.52 3.16 -10.77 22.49 24 485 The income group and regional classifications are those used by the World Bank
Table 4. Change in Debt Minus Deficit (δ)
μ (%) σ (%) Country Group
All Without Outliers* Overall Between
Min (%)
Max (%)
N. of countries
N. of observations
All Countries 4.93 3.15 21.84 13.29 -116.61 281.93 117 1872 By Region
By Income Groups Low 9.63 6.09 30.85 21.57 -116.61 247.90 34 440 Medium 4.87 3.09 21.88 8.87 -64.66 281.93 59 947 High 0.77 0.79 2.83 1.07 -12.16 14.07 24 485 The income group and regional classifications are those used by the World Bank. *Outliers are the top and bottom 2 percent of the distribution.
Countries Fixed Effects Country Country Country Country Country Ctry.-Year
DRER*LOW=DRER*MED 0.7654 0.7392 0.6757 0.6536 DRER*HIGH=DRER*MED 0.0612 0.0524 0.0396 0.0359 R-squared with country FE 0.4783 0.4825 0.4559 0.4584 0.4852 0.5025 Robust standard errors in parentheses. * Significant at 10 percent; ** significant at 5 percent; *** significant at 1 percent.
26
Table 11. The Determinants of δ (1) (2) (3) (4) INFLATION 24.443 24.541 26.064 24.646 (11.130)** (10.838)** (12.533)** (11.305)** DRER*LOW 15.872 15.998 (7.496)** (6.276)** DRER*MEDIUM 4.183 4.376 (5.526) (5.874) DRER*HIGH 35.377 35.300 (11.147)*** (10.440)*** DRER -0.493 -0.240 (1.814) (1.828) DEFAULT 2.091 2.338 (2.062) (1.860) BANKING CRISIS -2.902 -2.921 (2.519) (1.979) GDP GROWTH -0.324 -0.198 (0.118)*** (0.130) GOOD TIMES DUMMY -1.822 -1.582 (0.857)** (0.847)* BAD TIMES DUMMY 3.772 2.933 (1.241)*** (1.200)** Observations 1528 1529 1238 1529 Nr. of Countries 102 102 92 102 R-squared (within) 0.1064 0.1104 0.1670 0.1550 Fixed Effects Country Country Country Country Sample All Countries All Countries All Countries All Countries
All All All All All Sample Developing Countries Countries Countries Countries Countries Countries
Fixed Effects Country Country Country Country Country Ctry.-YearDRER: LOW=MED 0.7114 0.7447 0.681 0.6571 DRER: HIGH=MED 0.053 0.0514 0.0386 0.0349 R-squared with country FE 0.5074 0.5188 0.4939 0.4962 0.5213 0.5373 Robust standard errors in parentheses. * Significant at 10 percent; ** significant at 5 percent;
*** significant at 1 percent.
28
Table 13. Probit Regressions for Episodes in Top δ Decile
Table A2. Episodes with δ>10 Country Year Code Region Country Year Code Region Country Year Code Region
INDONESIA 1986 IDN EAP JAMAICA 2001 JAM LAC BURUNDI 1983 BDI SSA INDONESIA 1997 IDN EAP JAMAICA 1999 JAM LAC BURUNDI 2003 BDI SSA INDONESIA 1982 IDN EAP MEXICO 1987 MEX LAC BURUNDI 1992 BDI SSA INDONESIA 1978 IDN EAP MEXICO 1986 MEX LAC BURUNDI 1989 BDI SSA KOREA 1981 KOR EAP MEXICO 1994 MEX LAC CAMEROON 1994 CMR SSA MONGOLIA 1998 MNG EAP MEXICO 1982 MEX LAC CHAD 1999 TCD SSA MONGOLIA 1993 MNG EAP MEXICO 1989 MEX LAC CHAD 1995 TCD SSA MONGOLIA 1996 MNG EAP MEXICO 1985 MEX LAC CONGO, DEM. REP. OF 1989 ZAR SSA MONGOLIA 1994 MNG EAP NICARAGUA 1991 NIC LAC CONGO, DEM. REP. OF 1990 ZAR SSA PAPUA NEW GUINEA 1994 PNG EAP NICARAGUA 2001 NIC LAC CONGO, DEM. REP. OF 1997 ZAR SSA PAPUA NEW GUINEA 2001 PNG EAP NICARAGUA 2000 NIC LAC CONGO, DEM. REP. OF 1981 ZAR SSA PAPUA NEW GUINEA 1997 PNG EAP NICARAGUA 1995 NIC LAC CONGO, DEM. REP. OF 1993 ZAR SSA ALBANIA 1997 ALB ECA NICARAGUA 1998 NIC LAC CONGO, DEM. REP. OF 1992 ZAR SSA BELARUS 1994 BLR ECA NICARAGUA 1993 NIC LAC CONGO, DEM. REP. OF 1996 ZAR SSA BELARUS 1998 BLR ECA NICARAGUA 1992 NIC LAC CONGO, DEM. REP. OF 1994 ZAR SSA CROATIA 1998 HRV ECA NICARAGUA 1997 NIC LAC CONGO, DEM. REP. OF 1995 ZAR SSA CROATIA 1999 HRV ECA NICARAGUA 1999 NIC LAC CONGO, DEM. REP. OF 1980 ZAR SSA GEORGIA 1998 GEO ECA NICARAGUA 2002 NIC LAC COTE D IVOIRE 1995 CIV SSA GEORGIA 1999 GEO ECA NICARAGUA 1994 NIC LAC ETHIOPIA 1994 ETH SSA GEORGIA 1997 GEO ECA PANAMA 1993 PAN LAC ETHIOPIA 1993 ETH SSA HUNGARY 1993 HUN ECA PANAMA 1996 PAN LAC GABON 1991 GAB SSA RUSSIA 1998 RUS ECA PARAGUAY 2001 PRY LAC GHANA 1996 GHA SSA RUSSIA 1996 RUS ECA PERU 1991 PER LAC GUINEA 1998 GIN SSA RUSSIA 1995 RUS ECA PERU 1998 PER LAC KENYA 2000 KEN SSA RUSSIA 1994 RUS ECA PERU 1992 PER LAC LESOTHO 1996 LSO SSA RUSSIA 1999 RUS ECA PERU 1993 PER LAC LESOTHO 2000 LSO SSA SLOVAK REPUBLIC 2002 SVK ECA ST. VINCENT & GRENS. 1999 VCT LAC LESOTHO 1998 LSO SSA SLOVAK REPUBLIC 2001 SVK ECA BAHRAIN, KINGDOM OF 1988 BHR MNA LESOTHO 2001 LSO SSA TURKEY 1981 TUR ECA ISRAEL 1996 ISR MNA MALAWI 1986 MWI SSA TURKEY 2001 TUR ECA ISRAEL 1977 ISR MNA NIGERIA 1989 NGA SSA DENMARK 1993 DNK IND ISRAEL 1979 ISR MNA NIGERIA 1988 NGA SSA DENMARK 1983 DNK IND ISRAEL 1988 ISR MNA NIGERIA 1987 NGA SSA ICELAND 1984 ISL IND ISRAEL 1993 ISR MNA NIGERIA 1978 NGA SSA IRELAND 1983 IRL IND ISRAEL 1998 ISR MNA NIGERIA 1983 NGA SSA NORWAY 1986 NOR IND ISRAEL 1975 ISR MNA NIGERIA 1990 NGA SSA SWEDEN 1980 SWE IND ISRAEL 1985 ISR MNA NIGERIA 1981 NGA SSA ARGENTINA 2002 ARG LAC ISRAEL 1989 ISR MNA NIGERIA 1980 NGA SSA ARGENTINA 2003 ARG LAC ISRAEL 1981 ISR MNA NIGERIA 1993 NGA SSA BOLIVIA 1995 BOL LAC ISRAEL 1973 ISR MNA NIGERIA 1986 NGA SSA BOLIVIA 1993 BOL LAC ISRAEL 1974 ISR MNA RWANDA 1998 RWA SSA BRAZIL 1993 BRA LAC ISRAEL 1978 ISR MNA RWANDA 1994 RWA SSA BRAZIL 1992 BRA LAC ISRAEL 1984 ISR MNA RWANDA 2002 RWA SSA COSTA RICA 1991 CRI LAC ISRAEL 1980 ISR MNA RWANDA 2003 RWA SSA COSTA RICA 1998 CRI LAC ISRAEL 1986 ISR MNA RWANDA 1990 RWA SSA COSTA RICA 1978 CRI LAC ISRAEL 1990 ISR MNA RWANDA 1996 RWA SSA ECUADOR 1998 ECU LAC ISRAEL 1976 ISR MNA SENEGAL 1983 SEN SSA ECUADOR 1993 ECU LAC ISRAEL 1992 ISR MNA SIERRA LEONE 2003 SLE SSA ECUADOR 1999 ECU LAC ISRAEL 1987 ISR MNA SIERRA LEONE 1986 SLE SSA ECUADOR 1992 ECU LAC ISRAEL 1983 ISR MNA SIERRA LEONE 1992 SLE SSA EL SALVADOR 1987 SLV LAC ISRAEL 1982 ISR MNA SIERRA LEONE 1985 SLE SSA EL SALVADOR 1986 SLV LAC JORDAN 1988 JOR MNA SIERRA LEONE 1990 SLE SSA GUYANA 1995 GUY LAC JORDAN 1972 JOR MNA SIERRA LEONE 1988 SLE SSA GUYANA 1987 GUY LAC JORDAN 1990 JOR MNA SIERRA LEONE 1995 SLE SSA GUYANA 1989 GUY LAC LEBANON 1996 LBN MNA SIERRA LEONE 1999 SLE SSA GUYANA 1986 GUY LAC LEBANON 1994 LBN MNA SIERRA LEONE 1993 SLE SSA GUYANA 1994 GUY LAC LEBANON 1999 LBN MNA SIERRA LEONE 1989 SLE SSA GUYANA 1988 GUY LAC LEBANON 1993 LBN MNA SIERRA LEONE 1987 SLE SSA GUYANA 1980 GUY LAC MOROCCO 1983 MAR MNA SIERRA LEONE 1996 SLE SSA GUYANA 1976 GUY LAC MOROCCO 1997 MAR MNA SIERRA LEONE 1998 SLE SSA GUYANA 1982 GUY LAC MOROCCO 1992 MAR MNA SIERRA LEONE 1997 SLE SSA GUYANA 1979 GUY LAC SAUDI ARABIA 1996 SAU MNA SIERRA LEONE 2001 SLE SSA GUYANA 1991 GUY LAC SAUDI ARABIA 1998 SAU MNA SUDAN 1999 SDN SSA GUYANA 1985 GUY LAC MALDIVES 1985 MDV SAS SUDAN 1998 SDN SSA GUYANA 1975 GUY LAC MALDIVES 1982 MDV SAS SWAZILAND 1984 SWZ SSA GUYANA 1992 GUY LAC NEPAL 1991 NPL SAS UGANDA 2001 UGA SSA GUYANA 1990 GUY LAC PAKISTAN 1972 PAK SAS UGANDA 2002 UGA SSA HAITI 2002 HTI LAC SRI LANKA 1991 LKA SAS ZAMBIA 1993 ZMB SSA HONDURAS 1998 HND LAC SRI LANKA 1977 LKA SAS ZAMBIA 1982 ZMB SSA HONDURAS 1992 HND LAC SRI LANKA 1985 LKA SAS ZAMBIA 1990 ZMB SSA HONDURAS 1996 HND LAC BURUNDI 1996 BDI SSA ZAMBIA 1991 ZMB SSA HONDURAS 1993 HND LAC BURUNDI 1999 BDI SSA ZAMBIA 1995 ZMB SSA HONDURAS 1994 HND LAC BURUNDI 1998 BDI SSA ZAMBIA 1994 ZMB SSA HONDURAS 1990 HND LAC BURUNDI 1987 BDI SSA ZAMBIA 1996 ZMB SSA JAMAICA 1997 JAM LAC BURUNDI 2001 BDI SSA ZAMBIA 1986 ZMB SSA JAMAICA 1984 JAM LAC BURUNDI 1988 BDI SSA ZAMBIA 1998 ZMB SSA JAMAICA 1994 JAM LAC BURUNDI 1993 BDI SSA ZAMBIA 1984 ZMB SSA JAMAICA 1998 JAM LAC BURUNDI 1986 BDI SSA ZAMBIA 1985 ZMB SSA JAMAICA 1985 JAM LAC BURUNDI 1991 BDI SSA ZIMBABWE 1995 ZWE SSA JAMAICA 1983 JAM LAC BURUNDI 1995 BDI SSA JAMAICA 1993 JAM LAC BURUNDI 2002 BDI SSA
33
Table A3. Episodes with δ<-10
Country Year Code Region Country Year Code RegionINDONESIA 1998 IDN EAP SAUDI ARABIA 1999 SAU MNA ALBANIA 1998 ALB ECA MALDIVES 1984 MDV SAS AUSTRALIA 1980 AUS IND MALDIVES 1983 MDV SAS ECUADOR 2001 ECU LAC PAKISTAN 1973 PAK SAS ECUADOR 2000 ECU LAC CHAD 1994 TCD SSA GUYANA 1984 GUY LAC CHAD 1991 TCD SSA GUYANA 1996 GUY LAC CHAD 1998 TCD SSA GUYANA 1978 GUY LAC CONGO, DEM. REP. OF 1991 ZAR SSA HONDURAS 1991 HND LAC CONGO, REPUBLIC OF 2000 COG SSA JAMAICA 1992 JAM LAC COTE D IVOIRE 1998 CIV SSA NICARAGUA 1996 NIC LAC ETHIOPIA 1995 ETH SSA PANAMA 1989 PAN LAC GUINEA 1991 GIN SSA PANAMA 1990 PAN LAC LESOTHO 2003 LSO SSA ST. VINCENT & GRENS. 1997 VCT LAC LESOTHO 2002 LSO SSA SURINAME 1975 SUR LAC NIGERIA 1995 NGA SSA BAHRAIN, KINGDOM OF 1990 BHR MNA RWANDA 1995 RWA SSA BAHRAIN, KINGDOM OF 1987 BHR MNA SIERRA LEONE 2000 SLE SSA JORDAN 1992 JOR MNA SWAZILAND 1985 SWZ SSA JORDAN 1989 JOR MNA TOGO 1985 TGO SSA LEBANON 1997 LBN MNA UGANDA 1999 UGA SSA MOROCCO 1991 MAR MNA UGANDA 1992 UGA SSA OMAN 1992 OMN MNA ZAMBIA 1987 ZMB SSA OMAN 1993 OMN MNA ZIMBABWE 1996 ZWE SSA OMAN 1987 OMN MNA OMAN 1999 OMN MNA OMAN 1995 OMN MNA OMAN 1989 OMN MNA